On August 22, 2026, DeepSeek's official API pricing documentation added weekend billing adjustments: starting from midnight Beijing time on August 23, 2026, on Saturdays and Sundays, there will be no distinction between peak and valley periods, and call fees will be uniformly charged at the off-peak period. The rules for weekday peaks and valleys have not been abolished.
This change directly affects developers using DeepSeek's official API, especially for deferred batch generation, offline evaluation, data cleaning, and agent playback tasks. The lowest price is still half of the peak price, so scheduling non-emergency loads on weekends makes it easier to control costs than operating during weekday peak hours.
Which prices are calculated based on the weekend's low point
The three models listed in the official documentation all apply the new rules, including DeepSeek-v4-Flash, DeepSeek-v4-Pro, and the experimental DeepSeek-v4-Flash-Vision-Exp. For example, in RMB, the hit input for a million Flash tokens in the cache is 0.05 yuan at a low and 0.10 yuan at the peak; Cache missed input is 1.5 yuan and 3 yuan; The output is 4.5 yuan and 9 yuan. The lowest prices for the Pro are 0.15 yuan, 4.5 yuan, and 13.5 yuan respectively.
After the lowest price is unified over the weekend, the lower price range will be in effect throughout the day. However, the new regulation starts at midnight on Sunday, so the first effective date is August 23, 2026; From the following full weekend onward, Saturdays and Sundays will be held in the low slot.
On weekdays, Beijing time from 9:00 to 12:00 and 14:00 to 18:00 is still considered peak hours; on other working days, the off-peak prices are used. Model versions, context length, and concurrency limits have not changed due to this billing rule adjustment. If the caller does not distinguish between task priorities, the cost during the weekday peak will remain unchanged.
Another common misunderstanding is that the low price only changes the unit price and does not reduce actual token consumption. Long context, low cache hit rates, and overly long output can still drive up total costs. Cost control should focus on cache hits, input clipping, maximum output length, and failed retries rather than just moving execution times.
How can scheduling systems utilize the new rules?
For platforms with task queues, you can keep real-time conversations in the instant queue, and mark reviews, summary recalculations, and content indexes that do not require immediate returns as weekend batch tasks. Billing statistics should also be recorded separately by model, cache hit status, and input/output; otherwise, it is difficult to confirm whether savings come from price periods or token usage changes.
This adjustment from DeepSeek has added a stable weekend cost window for developers. It offers limited help to continuously online interactive products, but significantly improves budget certainty for scheduling calculations; Whether you truly save money ultimately depends on whether tasks can be safely postponed and whether the team has clear usage monitoring.
